Manifold Decoders: A Framework for Generative Modeling from Nonlinear Embeddings

High-dimensional data analysis and visualization constitute fundamental challenges in machine learning, where nonlinear dimensionality reduction (NLDR) techniques have proven instrumental in discovering low-dimensional embeddings that preserve essential structural properties of complex datasets. These methods, encompassing techniques such as t-distributed Stochastic Neighbor Embedding (t-SNE) [13], Isometric Mapping (Isomap) [12], Locally Linear Embedding (LLE) [10] and Laplacian Eigenmaps [1] excel at revealing intrinsic data manifolds and facilitating interpretable visualizations of high-dimensional phenomena. However, a critical architectural limitation pervades the entire class of traditional NLDR methods: they inherently lack reconstruction capabilities, operating as one-way transformations that map from high-dimensional input spaces to low-dimensional embeddings without providing mechanisms for inverse mapping. This fundamental asymmetry severely constrains the applicability of NLDR techniques in generative modelling, data synthesis, and interactive exploration scenarios where bidirectional transformations are essential. Unlike autoen-coders, which explicitly incorporate decoder architectures during training, classical manifold learning approaches such as t-SNE, Uniform Manifold Approximation and Projection (UMAP) [8], and diffusion maps optimize embeddings through eigen decomposition, neighbourhood preservation, or probabilistic formulations that do not naturally yield invertible mappings. Consequently, despite their superior performance in preserving local neighbourhood structures and global topological properties, these methods remain confined to analysis and visualization tasks. This work addresses the reconstruction gap in NLDR methods by developing specialized decoder architectures that enable bidirectional mapping between high-dimensional data and learned manifold representations.